Self-supervised 3D seismic data random noise suppression method based on 3D convolutional neural network

Abstract

The present invention belongs to the technical field of seismic data processing, and specifically relates to a method for suppressing random noise in 3D seismic data based on self-supervision of 3D convolutional neural network, comprising: 1. designing a 3D convolutional neural network architecture; 2. preparing 3D seismic data containing random noise; 3. designing self-supervision auxiliary tasks; 4. designing a strategy to avoid overfitting; 5. running the 3D convolutional neural network; 6. determining the optimal noise suppression result. Compared with the existing deep learning-based methods, this method solves the problem that the diversity of samples in the data set or the quality of labels limits the denoising and generalization capabilities of the neural network through a new type of 3D convolutional neural network assisted by a self-supervision strategy; considering the spatial correlation of effective signals in 3D seismic data, 3D convolution is used to overcome the limitations of 2D convolution in extracting 3D spatial structural features, improve the ability of the neural network to protect effective signals, and improve the signal-to-noise ratio of 3D seismic data.

Description

Technical Field

[0001] The present invention belongs to the technical field of seismic data processing, and in particular relates to a method for suppressing random noise in 3D seismic data based on a 3D convolutional neural network self-supervision. Background Art

[0002] In the actual seismic data acquisition process, due to many factors such as natural conditions and acquisition equipment, the effective signal received by the detector is usually mixed with the noise signal. As one of the main noises, random noise shows randomness in time and space, has a wide frequency band, and has no fixed propagation direction and speed. The existence of random noise increases the complexity of seismic data, reduces the signal-to-noise ratio, is not conducive to subsequent seismic data processing and interpretation, and does not meet the needs of high-precision seismic exploration.

[0003] Seismic random noise suppression methods can be mainly divided into: predictive filtering methods, modal decomposition methods, transform domain methods, inversion-based methods, and deep learning-based methods. Predictive filtering methods separate random noise and effective signals based on the predictability of the event axis in the spatial direction. Conventional fx domain predictive filtering is based on the linear event axis assumption. When the structure is complex, it will damage the effective signal. The modal decomposition method suppresses noise by eliminating the noise component from the multiple components decomposed by the seismic data. This type of method is difficult to achieve effective separation of effective signal components and noise components when the signal-to-noise ratio is low. It mainly includes empirical mode decomposition and variational mode decomposition. The transform domain method first transforms the seismic data to the corresponding domain, and designs a suitable threshold to separate the effective signal and random noise according to the difference in the properties of the effective signal and random noise in the transform domain. Finally, clean data is obtained through inverse transformation. This type of method is sensitive to the selection of thresholds, and mainly includes sparse transformation algorithms such as curvelet transform and wavelet transform. The basis function of traditional sparse transformation is generally fixed, so it cannot process seismic data with complex structures. Adaptive sparse transformation basis functions can be obtained through dictionary learning methods, such as the data-driven tight frame (DDTF) method. Although this method can handle complex data, its computational efficiency is low. Inversion-based random noise suppression methods directly invert effective signals from noise data, such as the Bayesian inversion method, but they are not suitable for seismic data with low signal-to-noise ratio or complex structure.

[0004] The seismic random noise suppression method based on deep learning mainly constructs a neural network model to mine the features of effective signals or random noise, so as to achieve effective separation of the two. It can be roughly divided into four categories: supervised learning, semi-supervised learning, self-supervised learning, and unsupervised learning. The supervised learning-based method requires training of a large amount of labeled data. The diversity of labeled data and the quality of labels limit the denoising ability of the neural network. The semi-supervised learning-based method alleviates the dependence of supervised learning-based methods on high-quality labeled data to a certain extent, but the ratio of labeled samples to unlabeled samples affects the performance of the neural network. The methods based on self-supervised learning and unsupervised learning directly mine implicit features from noise data without the need for effective signals as labels. The former requires the design of effective auxiliary tasks to automatically mine supervisory information from noise data, and the latter generally requires the use of network architecture as a prior model (such as the Deep Image Prior method) or relies on data preprocessing technology to improve performance.

[0005] Although the existing random noise suppression methods have their own advantages, they also have their own limitations in the application process. Traditional methods are prone to incomplete random noise separation or damage to effective signals when the assumptions are not met, the signal-to-noise ratio is low, the transform domain has overlap, the threshold selection is inappropriate, and the structure is relatively complex. Deep learning methods based on supervised learning and semi-supervised learning often need to use traditional random noise suppression methods when making labels, so it is unrealistic to obtain high-quality labels. When the test data and the training data have different data distributions, the generalization ability of these two types of methods is poor. Existing random noise suppression methods based on self-supervised learning and unsupervised learning have achieved some research results on two-dimensional seismic data, but there are relatively few studies on three-dimensional seismic data. The existing method (Quan, Y., Chen, M., Pang, T., et al. 2020. Self2Self With Dropout: Learning Self-Supervised Denoising From Single Image. 2020 IEEE / CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE.) applies Bernoulli sampling to a single 2D natural image, allowing the 2D convolutional neural network to achieve denoising in a self-supervised manner. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes a method for suppressing random noise in 3D seismic data based on 3D convolutional neural network self-supervision. This method only uses a single 3D noisy seismic data for training, avoiding the need for a large amount of high-quality labeled data, and solving the problem of sample diversity or label quality in the data set limiting the denoising and generalization capabilities of the neural network; 3D convolution is used to replace 2D convolution to extract the 3D spatial structural features of seismic data, thereby taking into account the spatial correlation of effective signals in 3D seismic data, solving the problem of insufficient 3D spatial structural features extracted by 2D convolution, improving the protection of effective signals, and improving the signal-to-noise ratio.

[0007] The technical solution adopted by the present invention is: a method for suppressing random noise of 3D seismic data based on 3D convolutional neural network self-supervision, the suppression method comprising the following steps:

[0008] Step 1: Design a 3D convolutional neural network architecture;

[0009] The 3D convolutional neural network is an Encoder-Decoder structure with jump connections, which consists of 3DPConv (3D partial convolution layer), 3DConv (3D convolution layer), LeakyReLU (leaky rectified linear unit), 3DMaxPooling (3D maximum pooling layer), 3DUpSampling (3D upsampling layer) and Dropout; wherein the 3DPConv (3D partial convolution layer) is obtained by replacing the 2D convolution filter in the 2D partial convolution layer with a 3D convolution filter, and 3DPConv can be expressed as: weight is the weight of the 3D convolution filter, bias is the corresponding bias, feature is the eigenvalue of the current 3D convolution sliding window, mask is the 3D binary mask, sum(1) is a matrix with the same dimension as mask but all elements are 1, Represents the multiplication of the elements at corresponding positions in the two matrices. The 3D partial convolution layer of the previous layer can automatically generate an updated mask for the 3D partial convolution layer of the next layer. The mask update function is

[0010] The module composed of 3DPConv-LeakyReLU is denoted as 3DPL, the module composed of 3DPConv-LeakyReLU-3DMaxPooling is denoted as 3DPLM, the module composed of Dropout-3DPConv-LeakyReLU is denoted as D3DPL, and the module composed of Dropout-3DConv-LeakyReLU is denoted as D3DCL;

[0011] The structure of Encoder is:

[0012]

[0013] The structure of Decoder is:

[0014]

[0015] The jump connection is to connect the feature vector of a certain stage of the encoder with the feature vector of a certain stage of the decoder in the channel direction. The jump connection mode of the network structure is:

[0016] The data input into the neural network is first mapped to the feature space using an encoder, and then an asymmetric decoder is used to reconstruct the effective signal in the seismic data and suppress the random noise. In this process, the features of the encoding stage are integrated into the decoding process with the help of jump connections, so that the decoder can obtain more fine-grained information, thereby better restoring the effective signal and improving the performance of suppressing random noise.

[0017] Step 2: Prepare 3D seismic data containing random noise;

[0018] The neural network designed in step 1 is used for denoising a single 3D noise data. In this way, there is no need to design a large amount of data for training, and only the data body that needs to suppress random noise is prepared; the 3D noise data d input to the neural network can be obtained by adding Gaussian random noise to a clean valid signal, and the clean valid signal is used as the true label c; the number of sampling points in the Xline direction is nx, the number of sampling points in the Inline direction is ni, and the number of sampling points in the Time direction is nt. Since the neural network can adapt to the dimension of the input data, nx, ni, and nt can be selected as any value greater than 1, but considering the training speed of the neural network, nx*ni*nt≤128 is preferred. 3 ;

[0019] Step 3: Design self-supervised auxiliary tasks;

[0020] Bernoulli sampling is used for 3D seismic data containing random noise, so that the 3D convolutional neural network can achieve self-supervised learning; the number of neural network iterations is N, and each training needs to generate a training data set by performing Bernoulli sampling on the 3D noise data d prepared in step 2 with probability p The input data is bs n represents an instance of a binary Bernoulli matrix, so the probability of each element in the 3D noise data d being set to 0 is 1-p, and the label is Example using Bernoulli sampling Train the neural network to construct supervision information in 3D seismic data and avoid identity mapping. The optimization goal of the neural network during each training is: Y θ Represents the parametric model of the neural network;

[0021] Step 4: Design a strategy to avoid overfitting;

[0022] The 3D convolutional neural network is used to suppress random noise in a single 3D seismic data. Fitting a single data volume with a large number of weight parameters can easily lead to overfitting. Designing a suitable strategy to avoid overfitting can effectively improve the performance of the neural network. The strategy to avoid overfitting is used for 3D convolutional neural networks and 3D seismic data containing random noise.

[0023] Strategy 1: Use the self-supervised auxiliary task designed in step 3 to train the 3D convolutional neural network designed in step 1, and test it after mI iterations, where m = 1, 2…M (MI ≤ N); in order to avoid overfitting, when testing after mI iterations, use the Dropout-based integration strategy to generate a K parameter model from the neural network model trained in this stage Use the same probability p as in step 3 to perform Bernoulli sampling on the 3D noise data d prepared in step 2 to generate the test data set As K parameter models The corresponding prediction result is Its mean That is the final test result after mI iterations;

[0024] Strategy 2: Since the input data of the 3D convolutional neural network is only a single 3D noise data d, in order to improve the robustness of the model and avoid overfitting, the training data set of step 3 is obtained by using Bernoulli sampling And the test data set of strategy 1 Before each sample in , randomly select whether to apply data enhancement technology to the 3D noise data d. If data enhancement technology is applied, randomly select one of the three methods: horizontal flip, vertical flip, and diagonal flip;

[0025] Step 5: Run the 3D convolutional neural network;

[0026] Select the optimizer, set the learning rate, set the initial mask matrix for each training and test, run the 3D convolutional neural network to suppress random noise, train and test the neural network at the same time, and get a valid signal after the training is completed The random noise suppressed is s m =do m ;

[0027] Step 6: Determine the optimal noise suppression result;

[0028] The signal-to-noise ratio is selected as the evaluation index to measure the ability to suppress random noise, and the formula is used Calculate m The signal-to-noise ratio between the known real effective signal c and the signal-to-noise ratio of the known real effective signal c is selected. m As the final result o*, the suppressed random noise s*=do*, and the local similarity graph between o* and s* is drawn to measure the protection capability of the effective signal.

[0029] Furthermore, in the step one, the number of filters of all 3DPConv is 32, the sliding step is (1,1,1), and the convolution kernel size is 3×3×3; the number of filters of 3DConv is 32, 32 and 1 respectively, the sliding step is (1,1,1), and the convolution kernel size is 1×3×3; the negative semi-axis slope of LeakyReLU is 0.1, the window size of each dimension of the input tensor of 3DMaxPooling is (2,2,2), and the sliding step is (2,2,2); the upsampling factor of 3DUpSampling is (2,2,2); the probability of each element in the input tensor of Dropout being set to zero is 0.15.

[0030] Beneficial effects of the present invention: A method for suppressing random noise in 3D seismic data based on 3D convolutional neural network self-supervision is proposed. The purpose is to solve the limitation of sample diversity or label quality in the data set on the denoising and generalization ability of the neural network through a new 3D convolutional neural network assisted by a self-supervision strategy; considering the spatial correlation of effective signals in 3D seismic data, 3D convolution is used to overcome the limitations of 2D convolution in extracting 3D spatial structural features, improve the protection ability of the neural network for effective signals, and improve the signal-to-noise ratio of 3D seismic data. Its main advantages are as follows:

[0031] (1) Design a new 3D convolutional neural network to extract the spatial structural characteristics of 3D seismic data, taking into account the spatial correlation of effective signals, improving the protection of effective signals, and improving the signal-to-noise ratio of 3D seismic data;

[0032] (2) A self-supervised strategy based on Bernoulli sampling combined with a 3D convolutional neural network is used to suppress random noise in single 3D seismic data. This method does not require labels to provide supervision information, saving the labor, computational, and economic costs of annotating large amounts of data. At the same time, it solves the problem that sample diversity or label quality in the data set affects the denoising and generalization capabilities of the neural network. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a flow chart of Embodiment 1;

[0034] Figure 2It is a diagram of the 3D convolutional neural network architecture in Example 1;

[0035] Figure 3 is a valid signal label diagram simulated in the first embodiment;

[0036] Figure 4 yes Figure 3 Random Gaussian noise map added in;

[0037] Figure 5 is the 3D synthetic data graph containing random noise in the first embodiment;

[0038] Figure 6 This is a graph showing the result of suppressing random noise using the method of the present invention in Example 1;

[0039] Figure 7 This is a graph showing the result of suppressing random noise using the DDTF method in Example 1;

[0040] Figure 8 is a random noise image suppressed by the method of the present invention in Example 1;

[0041] Fig. 9 is a random noise image suppressed by the DDTF method in Example 1;

[0042] Fig.10 yes Figure 6 and Figure 8 Local similarity graph between ;

[0043] Fig.11 yes Figure 7 and Fig. 9 The local similarity graph between . DETAILED DESCRIPTION

[0044] Embodiment 1

[0045] Referring to the figures, a method for suppressing random noise of 3D seismic data based on 3D convolutional neural network self-supervision is provided, the method comprising the following steps:

[0046] Step 1: Design a 3D convolutional neural network architecture;

[0047] First, configure the deep learning environment and install python3.7.0, numpy1.19.2, scipy1.4.1, matplotlib3.3.2, tensorflow-gpu2.6.0, and keras2.6.0;

[0048] In the created deep learning environment, use the application program interface provided by Tensorflow and Keras to build a 3D convolutional neural network with an Encoder-Decoder structure with skip connections, such as Figure 2 As shown, it specifically includes 3DPConv (3D partial convolution layer), 3DConv (3D convolution layer), LeakyReLU (leaky rectified linear unit), 3DMaxPooling (3D maximum pooling layer), 3DUpSampling (3D upsampling layer), and Dropout. The interface functions corresponding to each component are as follows:

[0049] 3DConv→tf.nn.conv3d; LeakyReLU→tf.nn.leaky_relu;

[0050] 3DMaxPooling→tf.nn.max_pool3d;

[0051] 3DUpSampling→tf.keras.layers.UpSampling3D;

[0052] Dropout → tf.nn.dropout. 3DPConv is obtained by replacing the 2D convolution filters in the 2D partial convolution layer with 3D convolution filters. 3DPConv can be expressed as: weight is the weight of the 3D convolution filter, bias is the corresponding bias, feature is the eigenvalue of the current 3D convolution sliding window, mask is the 3D binary mask, sum(1) is a matrix with the same dimension as mask but all elements are 1, Represents the multiplication of the elements at corresponding positions in the two matrices. The 3D partial convolution layer of the previous layer can automatically generate an updated mask for the 3D partial convolution layer of the next layer. The mask update function is According to the above principle, first import the module: from kerasimport backend as K, and then use K.conv3d and K.bias_add to implement 3DPConv.

[0053] The module consisting of 3DPConv—LeakyReLU is denoted as 3DPL, the module consisting of 3DPConv—LeakyReLU—3DMaxPooling is denoted as 3DPLM, the module consisting of Dropout—3DPConv—LeakyReLU is denoted as D3DPL, and the module consisting of Dropout—3DConv—LeakyReLU is denoted as D3DCL.

[0054] The structure of Encoder is:

[0055]

[0056] The structure of Decoder is:

[0057]

[0058] The jump connection is to connect the feature vector of a certain stage of the encoder with the feature vector of a certain stage of the decoder in the channel direction. The designed network structure jump connection method is as follows: Implemented using tf.concat.

[0059] The number of filters of all 3DPConv is 32, the sliding step is (1,1,1), and the convolution kernel size is 3×3×3; the number of filters of 3DConv is 32, 32 and 1 respectively, the sliding step is (1,1,1), and the convolution kernel size is 1×3×3; the negative semi-axis slope of LeakyReLU is 0.1, the window size of each dimension of the input tensor of 3DMaxPooling is (2,2,2), and the sliding step is (2,2,2); the upsampling factor of 3DUpSampling is (2,2,2); the probability of each element in the input tensor of Dropout being set to zero is 0.15.

[0060] Step 2: Prepare 3D seismic data containing random noise;

[0061] Use the suaddnoise command in the Seismic Unix open source program to simulate valid signals (such as Figure 3 )Add Gaussian random noise (such as Figure 4 ) to obtain 3D synthetic data containing noise (such as Figure 5 ), and use it as the 3D noise data d for the input neural network. The simulated valid signal is used as the true label c. Figure 3-5 The number of sampling points in the Xline direction of the 3D data volume shown is nx=96, the number of sampling points in the Inline direction is ni=96, and the number of sampling points in the Time direction is nt=96, nx*ni*nt≤128 3 .

[0062] Step 3: Design self-supervised auxiliary tasks;

[0063] Bernoulli sampling is used for 3D seismic data containing random noise, so that the 3D convolutional neural network can achieve self-supervised learning. The number of neural network iterations N is set to 50,000, and the 3D noise data d prepared in step 2 is sampled by Bernoulli with probability p = 0.75 to generate the training data set during each training. The input data is bs nrepresents an instance of a binary Bernoulli matrix. We can first generate a tensor whose elements are all 1 through tf.ones_like, and then use it as the input of tf.nn.dropout. The corresponding result is multiplied by the probability of each element being retained, 0.75, to obtain an instance of a binary Bernoulli matrix. Therefore, the probability of each element in d being set to 0 is 0.25, and the label is Example using Bernoulli sampling Train the neural network to construct supervision information in 3D seismic data and avoid identity mapping. The optimization goal of the neural network during each training is: Y θ Represents the parametric model of a neural network.

[0064] Step 4: Design a strategy to avoid overfitting;

[0065] The overfitting avoidance strategy designed in this method is used for 3D convolutional neural networks and 3D seismic data containing random noise;

[0066] Strategy 1: Use the self-supervised auxiliary task designed in step 3 to train the 3D convolutional neural network designed in step 1, set I to 1000, and test after 1000m iterations, where m = 1, 2…50 (M = 50, MI = 50000 ≤ N = 50000). In order to avoid overfitting, when testing after 1000m (m = 1, 2…50) iterations, a Dropout-based integration strategy is used to generate a K = 200 parameter model for the neural network model trained in this stage. Use the same probability p=0.75 as in step 3 to perform Bernoulli sampling on the 3D noise data d prepared in step 2 to generate the test data set As 200 parameter models The corresponding prediction result is Its mean That is the final test result after 1000m (m=1,2…50) iterations.

[0067] Strategy 2: Since the input data of the 3D convolutional neural network is only a single 3D noise data d, in order to improve the robustness of the model and avoid overfitting, the training data set of step 3 is obtained by using Bernoulli sampling And the test data set of strategy 1 Before each sample in , randomly choose whether to apply data augmentation technology to the 3D noise data d. If data augmentation technology is applied, randomly select one of the three methods: horizontal flip, vertical flip, and diagonal flip. Only calling tf.image.flip_left_right can achieve horizontal flip, only calling tf.image.flip_up_down can achieve vertical flip, and calling both can achieve diagonal flip.

[0068] Step 5: Run the 3D convolutional neural network;

[0069] Select the Adam optimizer, set the learning to 1e-4, and set the initial mask matrix for each training and test to the binary Bernoulli matrix that multiplies the corresponding elements of the 3D noise data d. Run the 3D convolutional neural network to suppress random noise. The neural network is trained and tested at the same time. After the training is completed, a valid signal is obtained. The random noise suppressed is s m =do m

[0070] Step 6: Determine the optimal noise suppression result;

[0071] Using the formula Calculate m The signal-to-noise ratio between the known real effective signal c, o 43 The signal-to-noise ratio between the real effective signal c is the highest. Therefore, the optimal result o*=o 43 ,like Figure 6 As shown, the result after noise suppression is consistent with the true label ( Figure 3 ) are almost the same, and the random noise is fully suppressed. The 3D noise data d( Figure 5 ) and the real effective signal c( Figure 3 ) is -1.37, that is Therefore, this method can effectively improve the signal-to-noise ratio of 3D seismic data without labels providing supervision information. The suppressed random noise s* = do*, such as Figure 8 As shown, the local similarity graph between o* and s* is as follows Fig.10 As shown, Fig.10 There is only a small amount of signal leakage. Figure 7 To suppress the Figure 5 The results obtained by the random noise in the data shown, the signal-to-noise ratio obtained by the DDTF method is 12.21. Compared with the DDTF method, this method has a better ability to suppress random noise, and the results obtained are ( Figure 6 ) is clearer. The random noise suppressed by the DDTF method is as follows Fig. 9As shown, Figure 7 and Fig. 9 The local similarity graph between the corresponding data is as follows Fig.11 shown. Fig.10 and Fig.11 compared to, Fig.11 There are many signal leaks in the simulation, which shows that this method has a strong ability to protect valid signals.

[0072] Therefore, the random noise suppression method for 3D seismic data based on self-supervision of 3D convolutional neural network provided by the present invention achieves effective suppression of random noise without relying on a large amount of labeled data to provide supervision information, improves the signal-to-noise ratio of 3D seismic data, and has a strong protection capability for effective signals.

Claims

1. A method for suppressing random noise in 3D seismic data based on 3D convolutional neural network self-supervision, characterized by: The pressing method comprises the following steps: Step 1: Design a 3D convolutional neural network architecture; The 3D convolutional neural network is an Encoder-Decoder structure with skip connections, which consists of 3DPConv, 3DConv, LeakyReLU, 3DMaxPooling, 3DUpSampling and Dropout; wherein 3DPConv is obtained by replacing the 2D convolution filter in the 2D partial convolution layer with a 3D convolution filter, and 3DPConv is expressed as: weight is the weight of the 3D convolution filter, bias is the corresponding bias, feature is the eigenvalue of the current 3D convolution sliding window, mask is the 3D binary mask, sum(1) is a matrix with the same dimension as mask but all elements are 1, It means that the elements at corresponding positions in the two matrices are multiplied. The 3DPConv of the previous layer automatically generates an updated mask for the 3DPConv of the next layer. The mask update function is The module composed of 3DPConv-LeakyReLU is denoted as 3DPL, the module composed of 3DPConv-LeakyReLU-3DMaxPooling is denoted as 3DPLM, the module composed of Dropout-3DPConv-LeakyReLU is denoted as D3DPL, and the module composed of Dropout-3DConv-LeakyReLU is denoted as D3DCL; The structure of Encoder is: Input&Mask —3DPL —3DPLM —3DPLM —3DPLM —3DPLM —3DPL The structure of Decoder is: 3DUpSampling①—D3DPL②—D3DPL③—3DUpSampling④—D3DPL⑤—D3DPL⑥—3DUpSampling⑦—D3DPL⑧—D3DPL⑨—3DUpSampling⑩—D3DPL —D3DPL —D3DCL —D3DCL —D3DCL —Output The jump connection is to connect the feature vector of a certain stage of the encoder with the feature vector of a certain stage of the decoder in the channel direction. The jump connection mode of this network architecture is: —①、 —④ —⑦ —⑩; The data input into the neural network is first mapped to the feature space using an encoder, and then an asymmetric decoder is used to reconstruct the effective signal in the seismic data and suppress the random noise. In this process, the features of the encoding stage are integrated into the decoding process with the help of jump connections. Step 2: Prepare 3D seismic data containing random noise; The neural network designed in step 1 is used for denoising a single 3D noise data. In this way, there is no need to design a large amount of data for training, and only the data body that needs to suppress random noise is prepared; the 3D noise data d input to the neural network is obtained by adding Gaussian random noise to the clean effective signal, and the clean effective signal is used as the true label c; the number of sampling points in the Xline direction is nx, the number of sampling points in the Inline direction is ni, and the number of sampling points in the Time direction is nt. Since the neural network can adapt to the dimension of the input data, nx, ni, and nt are selected as any value greater than 1; Step 3: Design self-supervised auxiliary tasks; Bernoulli sampling is used for 3D seismic data containing random noise, so that the 3D convolutional neural network can achieve self-supervised learning; the number of neural network iterations is N, and each training needs to generate a training data set by performing Bernoulli sampling on the 3D noise data d prepared in step 2 with probability p The input data is bs n represents an instance of a binary Bernoulli matrix, so the probability of each element in the 3D noise data d being set to 0 is 1-p, and the label is Example using Bernoulli sampling Train the neural network to construct supervision information in 3D seismic data and avoid identity mapping. The optimization goal of the neural network during each training is: Y θ Represents the parametric model of the neural network; Step 4: Design a strategy to avoid overfitting; Avoid overfitting strategies for 3D convolutional neural networks and 3D seismic data with random noise; Strategy 1: Use the self-supervised auxiliary task designed in step 3 to train the 3D convolutional neural network designed in step 1, and test it after mI iterations, where m = 1, 2...M. When testing after mI iterations, use the Dropout-based integration strategy to generate a K parameter model from the neural network model trained in this stage. Use the same probability p as in step 3 to perform Bernoulli sampling on the 3D noise data d prepared in step 2 to generate the test data set As K parameter models The corresponding prediction result is Its mean That is the final test result after mI iterations; Strategy 2: Since the input data of the 3D convolutional neural network is only a single 3D noise data d, the training data set of step 3 is obtained by using Bernoulli sampling And the test data set of strategy 1 Before each sample in , randomly select whether to apply data enhancement technology to the 3D noise data d. If data enhancement technology is applied, randomly select one of the three methods: horizontal flip, vertical flip, and diagonal flip; Step 5: Run the 3D convolutional neural network; Select the optimizer, set the learning rate, set the initial mask matrix for each training and test, run the 3D convolutional neural network to suppress random noise, train and test the neural network at the same time, and get a valid signal after the training is completed The random noise suppressed is s m =do m ; Step 6: Determine the optimal noise suppression result; The signal-to-noise ratio is selected as the evaluation index to measure the ability to suppress random noise, and the formula is used Calculate m The signal-to-noise ratio between the known real effective signal c and the signal-to-noise ratio of the known real effective signal c is selected. m As the final result o*, the suppressed random noise s*=do*, and the local similarity graph between o* and s* is drawn to measure the protection capability of the effective signal.

2. The method for suppressing random noise in 3D seismic data based on 3D convolutional neural network self-supervision according to claim 1, characterized in that: In step 1, the number of filters of all 3DPConv is 32, the sliding step size is (1, 1, 1), and the convolution kernel size is 3×3×3; The number of filters of 3DConv is 32, 32 and 1 respectively, the sliding step size is (1, 1, 1), and the convolution kernel size is 1×3×3; the negative semi-axis slope of LeakyReLU is 0.1, the window size of each dimension of the input tensor of 3DMaxPooling is (2,2,2), and the sliding step size is (2,2,2); the upsampling factor of 3DUpSampling is (2,2,2); the probability of each element in the input tensor of Dropout being set to zero is 0.15.

Images